State early warning method and system for strain clamp of power transmission line

Abstract

The invention provides a state early warning method and system for a strain clamp of a power transmission line, relates to the field of state monitoring of power equipment, and solves the technical problems of difficulty in data fusion and insufficient early warning accuracy in the prior art. The method comprises the following steps: collecting real-time data of a strain clamp through a multi-mode sensor array; the multi-mode sensor array comprises a thermosensitive resonator, an X-ray imaging device, an infrared temperature measurement sensor, a vibration sensor and an environment sensor; carrying out preprocessing and feature extraction on the collected real-time data to generate a multi-modal feature vector; inputting the multi-modal feature vector into a multi-modal interaction guidance network constructed based on a deep learning algorithm, and performing mutual guidance and feature sharing among multi-category early warning tasks through a task interaction guidance module; and based on the output of the multi-modal interaction guidance network, calculating the early warning probability of the multi-class tasks, and performing hierarchical early warning decision.

Description

Technical Field

[0001] This invention belongs to the field of power equipment condition monitoring, specifically a condition early warning method and system for tension clamps of transmission lines. Background Technology

[0002] Tension clamps, as key hardware for overhead transmission lines, directly affect the safety and stability of the power grid through their crimping quality and operational status. Currently, methods for monitoring the condition of tension clamps have significant limitations: Firstly, existing technologies often rely on single sensors or simple combinations of a few sensors, making it difficult to comprehensively capture the complex operating conditions of the clamps. Secondly, the ability to fuse and process multi-source heterogeneous data is weak; most methods use linear or nonlinear mathematical models for condition scoring, making it difficult to uncover deep correlations and spatiotemporal dependencies between data. Such models have poor adaptability to complex fault modes, low early warning accuracy, and frequent false alarms and missed alarms. Furthermore, existing methods are mostly based on threshold judgments or periodic inspections, lacking real-time and proactive capabilities, and unable to achieve early risk warnings. Summary of the Invention

[0003] This application provides a method and system for early warning of the condition of tension clamps in transmission lines, which solves the technical problems of difficult data fusion and insufficient early warning accuracy in the prior art.

[0004] To achieve the above objectives, this application adopts the following technical solution: Firstly, a method for early warning of the condition of tension clamps in transmission lines is provided, including: Real-time data of tension clamps are acquired through a multimodal sensor array; the multimodal sensor array includes a thermal resonator, an X-ray imaging device, an infrared temperature sensor, a vibration sensor, and an environmental sensor; the thermal resonator includes a resonant cavity, an excitation unit, a detection unit, and a signal processing circuit. The collected real-time data is preprocessed and features are extracted to generate multimodal feature vectors; The multimodal feature vectors are input into a multimodal interactive guidance network constructed based on deep learning algorithms, and mutual guidance and feature sharing among multiple types of early warning tasks are achieved through the task interaction guidance module; Based on the output of the multimodal interactive guidance network, the early warning probability of multiple types of tasks is calculated, and hierarchical early warning decisions are made.

[0005] Based on the above technical solutions, the condition early warning method for tension clamps in transmission lines provided in this application achieves comprehensive and high-precision real-time monitoring of the tension clamp condition through the integrated application of a multi-modal sensor array. Utilizing a deep learning-based multi-modal interactive guidance network, the accuracy and synergy of early warnings for multiple categories such as defects, overheating, and corrosion are significantly improved through mutual guidance and feature sharing among tasks. Simultaneously, combining preprocessing and feature extraction techniques transforms heterogeneous data into a unified multi-modal feature vector, optimizing data processing efficiency. Finally, through a hierarchical early warning decision-making mechanism, refined risk identification and response are achieved, thereby significantly improving the safety and reliability of transmission line operation and reducing accident risks.

[0006] Furthermore, the thermistor operates based on a temperature-frequency conversion mechanism, including the following steps: The surface temperature of the tension clamp is sensed by the thermistor material in the resonant cavity, and the frequency shift caused by the surface temperature change is obtained by the excitation unit and the detection unit. The frequency offset is converted into a temperature value based on a pre-calibrated thermistor model, which is: Δf = k × ΔT + The calculation is performed, where Δf is the frequency shift, ΔT is the temperature change, and k is the thermistor coefficient. This is an environmental compensation item; The processed temperature value is sent to the edge computing node via a wireless transmission module.

[0007] Furthermore, the real-time data includes: Time-series data of surface temperature of tension clamp acquired using a thermistor; X-ray image data acquired using an X-ray imaging device; Temperature time-series data collected using an infrared temperature sensor; Vibration signal sequence data acquired using vibration sensors; Environmental data such as temperature, humidity, wind speed, and rainfall are collected using environmental sensors.

[0008] Furthermore, the preprocessing and feature extraction of the acquired real-time data to generate multimodal feature vectors includes: The X-ray image is enhanced using the Retinex algorithm and combined with a generative adversarial network for super-resolution reconstruction, outputting the first feature vector; A one-dimensional convolutional network is used to extract temporal features from the temperature time-series data and vibration signal sequence data to obtain a second feature vector; the temporal features include, but are not limited to, heating rate, fluctuation mode and frequency domain energy. The environmental data is embedded and represented through a fully connected layer to obtain the third feature vector; The first feature vector, the second feature vector, and the third feature vector are concatenated to obtain the fused multimodal feature vector.

[0009] Furthermore, the multimodal interaction guidance network includes a shared feature extraction module, a task interaction guidance module, and an early warning output module; wherein, The shared feature extraction module includes a cross-modal fusion unit and a feature enhancement unit, which are used to perform cross-modal alignment and feature optimization on multimodal feature vectors to obtain shared feature vectors. The cross-modal alignment means calculating the spatiotemporal correlation between different modal features through a multi-head attention mechanism. The feature optimization means using residual connection and batch normalization techniques to denoise and unify the scale of the fused features, and assigning higher weights to high-confidence modal features through an adaptive weight adjustment mechanism. The task interaction guidance module includes a task attention gate and an interactive task guide, which are used to perform weight allocation and feature guidance on the shared feature vector among multiple tasks to obtain specific feature vectors for each task. The weight allocation means calculating attention weights based on the correlation between tasks, and the feature guidance means realizing the transfer and correction of features between tasks through a gated loop unit, using the feature vectors output by different tasks as prior inputs for other tasks, and cross-guiding feature sharing and learning between tasks. The warning output module includes multiple warning classification heads, which are used to perform multi-class probability calculations on the specific feature vectors of each task to obtain the warning probability of the multi-class task.

[0010] Furthermore, the internal workflow of the shared feature extraction module is as follows: The cross-modal fusion unit receives multimodal feature vectors, calculates the correlation weights between different modalities through a multi-head attention mechanism, and performs dimensional alignment of each modal feature through an adaptive fully connected layer to obtain dimensional aligned modal features. Based on the correlation weights and the confidence of each modality, the modal features after dimension alignment are weighted and fused to obtain cross-modal fused features; The cross-modal fusion features are input into the feature enhancement unit, where the feature expression is strengthened and gradient vanishing is alleviated through the residual connection structure. The fusion features are then denoised using batch normalization techniques to obtain the denoised fusion features. The fused features after noise reduction are scaled by a fully connected layer, and a shared feature vector is output.

[0011] Furthermore, the confidence level of each modality is used to quantify the reliability and validity of each type of modality data in the modal feature vector, and the acquisition method includes: For each image modality, the Laplacian operator is used to calculate the gradient magnitude at the image edges, and the percentage of pixels with gradient magnitudes greater than a preset threshold is counted to obtain an image sharpness score. For time-series modes of temperature sequences and vibration signal characteristics, the effective signal and noise components in the time-series data are separated, and the signal-to-noise ratio (SNR) score is obtained according to the SNR calculation formula: SNR = effective signal amplitude squared / noise signal amplitude squared. For scalar modes of environmental data, the stability of the scalar data is calculated using the formula of volatility coefficient = standard deviation / mean, and then converted into a stability score by "1 - volatility coefficient"; The image sharpness score, signal-to-noise ratio score, and stability score are all processed by Min-Max normalization, and each type of score is mapped to the [0,1] interval to obtain the confidence level corresponding to each mode.

[0012] Furthermore, the internal workflow of the task interaction guidance module is as follows: The shared feature vector is received, and a dedicated query vector is generated for each of the multiple categories of early warning tasks through a preset independent fully connected layer in the task attention gate. Based on the specific query vector, the correlation of multiple categories is calculated using cosine similarity, and the calculation results are normalized using Softmax to obtain the task attention weight matrix. The shared feature vectors are weighted and summed based on the task attention weight matrix to obtain the initial task feature vectors for each task. The initial task feature vectors are sorted according to a preset logic and then input into the interactive task guide. The feature vectors of the preceding tasks are used as the prior inputs of the subsequent tasks through a series of gated loop units. The task key features are transmitted and corrected to obtain the intermediate guidance features of each task. For each guided feature, the guided intermediate feature is fused with the guided intermediate features of other tasks through a feature cross-concatenation operation. Then, the dimensionality is compressed through a fully connected layer and the feature expression is enhanced by combining the GELU activation function, and the specific feature vector of each task is output.

[0013] Furthermore, the internal workflow of the early warning output module is as follows: Each task-specific feature vector is received by the task interaction guidance module through the warning classification head corresponding to each warning task. Each warning classification head enhances and compresses the specific feature vector input through a fully connected layer and the GELU activation function, performs regularization by combining a Dropout layer, and then calculates the initial probability of the corresponding warning task through the Sigmoid activation function. The initial probability of multiple categories of early warning tasks is received through the security classification head, and the initial probability of the security status category is calculated based on the reverse correlation logic through a fully connected layer and a Sigmoid activation function. The initial probabilities of all warning categories and the initial probabilities of the safety status category are normalized using Softmax so that the sum of the probabilities of each category is 1. Finally, the probability distribution of the multi-category warning task and the safety status category is output.

[0014] Furthermore, the reverse correlation logic refers to the negative correlation between the initial probability of the safety status category and the initial probability of each early warning task. By quantifying the comprehensive anomaly degree of multiple categories of early warning tasks, the confidence level of the tension clamp being in the safety status category is derived in reverse. The operation steps include: The initial probabilities of the multi-category early warning tasks are comprehensively calculated, and the anomaly index is obtained through a weighted summation formula; The anomaly index is reverse-mapped, and the basic safety confidence value is obtained through a linear transformation of "1-anomaly index". The basic security confidence values ​​are sequentially input into the fully connected layer and the Sigmoid function to obtain the initial probability of the security state category.

[0015] Furthermore, the multi-category early warning tasks and the preset logical order are: defect early warning, overheating early warning, corrosion early warning, and degradation early warning.

[0016] Secondly, this application provides a status early warning system for tension clamps in transmission lines, comprising: a data acquisition module, an intelligent processing module, and an early warning decision module; wherein, The data acquisition module is used to acquire real-time data of the tension clamp through a multi-modal sensor array. The multi-modal sensor array includes a thermal resonator, an X-ray imaging device, an infrared temperature sensor, a vibration sensor, and an environmental sensor. The thermal resonator includes a resonant cavity, an excitation unit, a detection unit, and a signal processing circuit. The intelligent processing module is used to preprocess and extract features from the collected real-time data, generate multimodal feature vectors, and input the multimodal feature vectors into a multimodal interactive guidance network built based on deep learning algorithms. Through the task interactive guidance module, mutual guidance and feature sharing are carried out between multiple categories of early warning tasks. The early warning decision module is used to calculate the early warning probability of multiple types of tasks based on the output of the multimodal interactive guidance network, and to make hierarchical early warning decisions.

[0017] Compared with the prior art, the beneficial effects of this application are: This application achieves high-precision monitoring of the surface temperature of tension clamps using a thermistor. Its temperature-frequency conversion mechanism effectively avoids errors caused by environmental interference in traditional infrared thermometry. Combined with dynamic correction of the environmental compensation term, data reliability is significantly improved. The collaborative operation of a multimodal sensor array enables comprehensive perception of the clamp's internal structure, temperature distribution, mechanical state, and external environment, overcoming the limitations of single-sensor monitoring. Advanced preprocessing techniques such as the Retinex algorithm, generative adversarial networks, and one-dimensional convolutional networks are used to transform heterogeneous data into unified multimodal feature vectors, providing high-quality input for subsequent analysis.

[0018] This application proposes an intelligent collaborative mechanism based on a multimodal interactive guidance network. The shared feature extraction module achieves cross-modal spatiotemporal alignment through a multi-head attention mechanism and dynamically optimizes feature weights by combining confidence evaluation, enhancing the robustness of data fusion. The task interaction guidance module utilizes task attention gates and gated recurrent units to achieve feature transfer and cross-guidance between multiple tasks such as defects, overheating, corrosion, and degradation, enabling mutual promotion of early warning tasks and reducing false alarms and missed alarms. The early warning output module quantifies the comprehensive anomaly degree through reverse correlation logic and outputs probabilistic decision results, improving the accuracy and interpretability of early warnings. This end-to-end intelligent processing flow can improve the efficiency and accuracy of state recognition.

[0019] In summary, this application, through its integrated design of multimodal perception, intelligent fusion, and adaptive decision-making, not only solves the problems of inaccurate monitoring and delayed early warning in complex environments using traditional methods, but also achieves collaborative optimization of multiple types of early warning tasks, providing efficient and reliable technical support for the safe operation of transmission lines. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 A system architecture diagram for a condition early warning system for tension clamps of transmission lines is provided in this application embodiment; Figure 2 A flowchart illustrating a status early warning method for tension clamps in transmission lines, provided as an embodiment of this application; Figure 3 A flowchart illustrating another method for early warning of the condition of tension clamps for transmission lines provided in this application embodiment; Figure 4A flowchart illustrating another method for early warning of the condition of tension clamps for transmission lines provided in this application embodiment; Figure 5 This is a flowchart illustrating a method for early warning of the condition of tension clamps for power transmission lines, provided in an embodiment of this application. Detailed Implementation

[0022] In the description of this application, unless otherwise stated, " / " means "or," for example, A / B can mean A or B. The "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone. Furthermore, "at least one" means one or more, and "multiple" means two or more. The terms "first," "second," etc., do not limit the quantity or order of execution, and "first," "second," etc., do not necessarily imply differences.

[0023] It should be noted that, in this application, the terms "exemplary" or "for example" are used to indicate that something is being described as an example, illustration, or illustration. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or design solutions. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0024] The method for early warning of the condition of tension clamps in transmission lines provided in this application embodiment can be applied to, for example... Figure 1 The above describes a condition warning system for tension clamps in transmission lines, such as... Figure 1 As shown, the system includes: a data acquisition module, an intelligent processing module, and an early warning and decision-making module; among which, The data acquisition module is used to acquire real-time data of the tension clamp through a multi-modal sensor array. The multi-modal sensor array includes a thermal resonator, an X-ray imaging device, an infrared temperature sensor, a vibration sensor, and an environmental sensor. The thermal resonator includes a resonant cavity, an excitation unit, a detection unit, and a signal processing circuit. The intelligent processing module is used to preprocess and extract features from the collected real-time data, generate multimodal feature vectors, and input the multimodal feature vectors into a multimodal interactive guidance network built based on deep learning algorithms. Through the task interaction guidance module, mutual guidance and feature sharing are carried out between multiple categories of early warning tasks. The early warning decision module is used to calculate the early warning probability of multiple types of tasks based on the output of the multimodal interactive guidance network, and to make hierarchical early warning decisions.

[0025] To address the technical problems of data fusion difficulties, insufficient early warning accuracy, limitations of single sensor monitoring, poor real-time performance, and weak adaptability to complex faults in existing technologies, this application provides a method and system for early warning of the status of tension clamps in transmission lines. The method includes: Real-time data of tension clamps are acquired through a multi-modal sensor array; the multi-modal sensor array includes a thermal resonator, an X-ray imaging device, an infrared temperature sensor, a vibration sensor, and an environmental sensor; the thermal resonator includes a resonant cavity, an excitation unit, a detection unit, and a signal processing circuit. The collected real-time data is preprocessed and features are extracted to generate multimodal feature vectors; Multimodal feature vectors are input into a multimodal interactive guidance network built based on deep learning algorithms, and mutual guidance and feature sharing among multiple types of early warning tasks are achieved through the task interaction guidance module; Based on the output of the multimodal interactive guidance network, the early warning probability of multiple types of tasks is calculated, and hierarchical early warning decisions are made.

[0026] Based on this, the method can achieve comprehensive monitoring and accurate early warning of the tension clamp status, thereby improving the safety and reliability of transmission line operation.

[0027] like Figure 2 As shown in the embodiment of this application, a method for early warning of the condition of tension clamps for transmission lines includes: S1. Real-time data of tension clamps are collected through a multi-modal sensor array.

[0028] Among them, multimodal sensors include, but are not limited to, thermistors, X-ray imaging devices, infrared temperature sensors, vibration sensors, and environmental sensors, and may also include ultrasonic sensors, ultraviolet imaging sensors, partial discharge sensors, etc.

[0029] In some implementation methods, the deployment of sensors can be flexibly selected according to the monitoring scenario, including fixed installation on the tension clamp body and surrounding components, mobile inspection and data collection via drones, or fixed-point supplementary data collection using handheld portable devices.

[0030] It should be noted that the core function of collecting real-time data is to break through the monitoring limitations of a single sensor, obtain the structural status, operating parameters and external environmental influencing factors of the tension clamp from different dimensions, and provide a comprehensive data source for subsequent analysis.

[0031] For example, a thermistor and an infrared temperature sensor are installed at the crimping part of the tension clamp to collect surface temperature data simultaneously; a vibration sensor is deployed at the conductor connection end to capture mechanical vibration signals during operation; and an environmental sensor is set up near the monitoring area to record external environmental parameters such as temperature and humidity.

[0032] S2. Preprocess and extract features from the collected real-time data to generate multimodal feature vectors.

[0033] Preprocessing includes, but is not limited to, data denoising, data normalization, missing value completion, and outlier removal; feature extraction can use Sobel edge detection and histogram equalization to extract morphological and texture features from image data, sliding window statistics, Fourier transform, and wavelet transform to extract trend and frequency features from time series data, and mean calculation, variance analysis, and extreme value extraction to extract statistical features from environmental scalar data.

[0034] In some implementations, multimodal feature vectors refer to a unified dimensional vector formed by integrating different types of features such as image features, time-series features, and environmental features. The generation methods include feature concatenation, weighted fusion, and feature concatenation, with the aim of transforming heterogeneous data into standardized inputs that can be used for model computation.

[0035] It should be noted that the core function of preprocessing is to purify data quality and reduce the interference of noise and outliers on subsequent analysis; feature extraction, on the other hand, is to extract key and effective information from the raw data, reduce data dimensionality, and improve the efficiency of subsequent model processing.

[0036] For example, the X-ray imaging data is first denoised by median filtering, and then the edge features of the internal structure of the clamp are extracted using the Sobel operator; the average heating rate every 5 minutes is calculated for the temperature time series data by sliding window, and then the frequency domain features are extracted by Fourier transform; the mean and variance of the environmental humidity and wind speed data are calculated over 24 hours, and finally the three types of features are concatenated in dimensional order to generate a multimodal feature vector.

[0037] S3. Input the multimodal feature vectors into the multimodal interactive guidance network built based on deep learning algorithms, and use the task interaction guidance module to guide each other and share features among multiple types of early warning tasks.

[0038] Among them, the task interaction guidance module refers to the functional module that can realize information interaction and collaborative optimization among multiple types of early warning tasks. It can usually be constructed based on methods such as multi-head attention mechanism, graph neural network, gated recurrent unit, and mutual information maximization criterion.

[0039] In some implementations, in the field of deep learning, mutual guidance between features can be achieved through feature distillation, attention weight allocation, cross-task feature projection, etc., by strengthening effective features and suppressing redundant features to achieve positive guidance between tasks; feature sharing refers to multiple early warning tasks sharing the same set of basic feature extraction layers or some intermediate features to avoid repeated calculations, while allowing different tasks to learn from each other's effective information during the learning process.

[0040] It should be noted that the core function of this step is to uncover the deep correlation between different modal features and different early warning tasks, and to improve the overall performance of multiple early warning tasks through collaborative learning, thereby avoiding the limitations of single-task learning.

[0041] For example, a task interaction guidance module based on a multi-head attention mechanism is constructed. First, the correlation weights between tasks such as defect warning and overheating warning are calculated, and then the weights are assigned to the corresponding shared features, so that highly correlated tasks can share effective features first. At the same time, the temperature-related features learned by the overheating warning task are passed to the corrosion warning task through a gated loop unit, providing prior information on the temperature influence for corrosion warning.

[0042] S4. Based on the output of the multimodal interactive guidance network, calculate the early warning probability of multiple types of tasks and make hierarchical early warning decisions.

[0043] Among them, the output of the multimodal interactive guidance network is the feature representation corresponding to each early warning task. These features can accurately characterize the state differences of tension wire clamps in the corresponding dimensions, and can provide a basis for quantitative analysis for graded early warning.

[0044] In some implementations, the probability of early warning can be calculated using classification algorithms such as logistic regression, softmax regression, and support vector machines, mapping the network output features to probability values ​​corresponding to each category; the hierarchical early warning decision can set multiple thresholds according to actual application needs, or use methods such as fuzzy comprehensive evaluation and hierarchical analysis to transform the probability values ​​into different levels of early warning results such as safety, general warning, severe warning, and emergency response.

[0045] It should be noted that the core function of tiered early warning decision-making is to achieve refined risk response, enabling operation and maintenance personnel to take targeted measures according to the early warning level, thus balancing early warning sensitivity and response efficiency.

[0046] For example, the probability values ​​of four types of early warning tasks—defect, overheating, corrosion, and degradation—are calculated using Softmax regression. A probability value of 0-0.3 is set as the safety level, 0.3-0.6 as the general early warning level, 0.6-0.8 as the severe early warning level, and above 0.8 as the emergency response level. The corresponding early warning level and response suggestions are output based on the probability value of each task.

[0047] Based on the above technical solution, the present application provides a condition early warning method for tension clamps in transmission lines. This method uses a multi-modal sensor array to comprehensively collect multi-dimensional condition data of the tension clamps, generates high-quality multi-modal feature vectors through flexible and adaptable preprocessing and feature extraction methods, enhances multi-task collaborative performance through a multi-modal interactive guidance network, and achieves precise risk response through scientific and reasonable hierarchical early warning decision-making. This significantly improves the accuracy, real-time performance, and comprehensiveness of early warning, effectively reduces accident risks, and ensures the safe and stable operation of transmission lines.

[0048] In one possible implementation of this application embodiment, the above-mentioned S1 can be specifically implemented by the following S101, S102 and S103, which are described in detail below: S101. Collect and process the surface temperature data of the tension clamp through a thermistor.

[0049] Among them, the thermistor is the core component in the multi-mode sensor array used for high-precision monitoring of the surface temperature of the clamp. Its operation relies on the temperature-frequency conversion mechanism, which can convert temperature changes into quantifiable frequency signals, thereby achieving accurate temperature measurement.

[0050] In some implementations, the specific workflow of a thermistor includes the following key steps: The first step is to sense the surface temperature of the tension clamp through the thermistor in the resonant cavity, and at the same time use the excitation unit to generate an excitation signal and the detection unit to capture the frequency shift caused by the temperature change. The second step involves converting the frequency offset into a temperature value based on a pre-calibrated thermistor model. The calculation formula for this thermistor model is Δf = k × ΔT + Δf represents the frequency offset (unit: Hz), ΔT represents the temperature change (unit: ℃), and k represents the thermistor coefficient (unit: Hz / ℃, which is determined by the characteristics of the thermistor material and needs to be calibrated in advance through experiments). This represents the environmental compensation item, used to correct the interference of ambient temperature and humidity on the frequency signal, and is dynamically adjusted according to the actual monitoring environment. The third step is to send the processed temperature value to the edge computing node via a wireless transmission module to ensure that the data is uploaded in real time.

[0051] It should be noted that the temperature-frequency conversion mechanism of the thermistor can effectively avoid the problem of traditional infrared thermometry being easily interfered with by environmental factors such as dust and fog. Combined with the dynamic correction of the environmental compensation term, it can significantly improve the reliability and accuracy of temperature data, providing high-quality basic data for subsequent overheating early warning.

[0052] S102. Collect structural, vibration and environmental data of the tension clamp using other types of sensors.

[0053] Other types of sensors include X-ray imaging devices, infrared temperature sensors, vibration sensors, and environmental sensors, which supplement the state information of the tension clamp from different dimensions and complement the temperature data of the thermistor.

[0054] In some implementations, the data collected by various sensors and the collection targets are as follows: The X-ray imaging device is used to acquire X-ray image data inside the tension clamp, focusing on capturing structural defects such as gaps and broken strands in the crimping area; Infrared temperature sensors are used to collect time-series temperature data from different areas of the wire clamp, which is then cross-validated with the temperature data from the thermistor. Vibration sensors are used to collect vibration signal sequence data of wire clamps during operation, reflecting the stability of the mechanical connection; Environmental sensors are used to collect environmental data such as temperature, humidity, wind speed and rainfall in the monitored area, and to analyze the impact of the external environment on the operating status of tension clamps.

[0055] It should be noted that the collaborative acquisition of multiple types of sensors is the key to breaking the limitations of single sensor monitoring; among them, X-ray imaging devices focus on internal structure, temperature sensors focus on thermal state, vibration sensors focus on mechanical state, and environmental sensors focus on external influences. The combination of these four types of data can achieve comprehensive perception of the operating status of tension clamps.

[0056] S103. Summarize and preliminarily label the real-time data collected by various sensors.

[0057] The core of this step is to integrate the scattered data collected by various sensors in S101 and S102 into the same data receiving end, namely the edge computing node, and add metadata tags to provide a structured data foundation for subsequent preprocessing stages.

[0058] In some implementations, the specific operations for data aggregation and tagging include: Edge computing nodes can simultaneously receive temperature values ​​from the thermal resonator, X-ray image data, infrared temperature time-series data, vibration signal sequences, and environmental data through multiple interfaces; Add a uniform metadata tag to each type of data. The tag content includes the acquisition time, acquisition sensor number, corresponding tension clamp device number and installation location. Store the tagged data in the local database according to the directory structure of "sensor type-data type-acquisition time".

[0059] Based on the above technical solution, S1 achieves full-dimensional, high-precision, and structured acquisition of real-time data of tension clamps through high-precision temperature measurement with a thermistor, collaborative acquisition by multiple types of sensors, and data aggregation and labeling. Among them, the temperature-frequency conversion mechanism of the thermistor solves the accuracy problem of traditional temperature measurement, the division of labor among multiple sensors covers the full dimensions of "structure-thermal-mechanical-environment", and the data aggregation and labeling lays a structured foundation for subsequent preprocessing and feature extraction, ultimately providing high-quality data source support for improving the accuracy of early warning.

[0060] In one possible implementation of the embodiments of this application, combined with Figure 2 ,like Figure 3 As shown, the above S2 can be implemented through the following S201, S202 and S203, which are explained in detail below: S201. Preprocess and extract features from the X-ray image data to generate the first feature vector.

[0061] Among them, X-ray image data is the core data reflecting the internal structural state of tension clamps, such as possible crimping gaps, broken strands, and foreign matter inclusions.

[0062] In some implementations, image data preprocessing and feature extraction include the following steps: First, the Retinex algorithm is used to enhance the X-ray image. By separating the illuminance component and the reflection component of the image, the influence of uneven ambient lighting on the image is suppressed, and the detail contrast of the internal structure of the clamp is enhanced, such as the grayscale difference of the crimping interface. Next, super-resolution reconstruction is performed by combining a Generative Adversarial Network (GAN). Through adversarial training between the generator and discriminator in the GAN, the low-resolution blurred image is reconstructed into a high-resolution image, solving the problem of detail loss in X-ray imaging due to equipment precision or distance. Then, feature extraction is performed on the reconstructed high-resolution image. Features such as structural edges and texture distribution can be extracted through convolution operations and finally integrated into a fixed-dimensional vector, namely the first feature vector.

[0063] In addition, to quantify the reliability of image data, an image sharpness score needs to be calculated: the edge gradient magnitude of each pixel in the image is calculated using the Laplacian operator, and the percentage of pixels with gradient magnitude greater than a preset threshold (such as 50) is counted. This percentage is the image sharpness score. The higher the score, the richer the image details and the stronger the data reliability.

[0064] It should be noted that the Retinex algorithm solves the problem of low image contrast, and GAN super-resolution reconstruction solves the problem of blurred details. The combination of the two can clearly reveal the tiny defects inside the clamp, laying the foundation for the accuracy of subsequent feature extraction.

[0065] S202. Preprocess and extract features from time-series data and environmental data to generate second and third feature vectors.

[0066] The time-series data includes temperature time-series data collected by the thermal resonator and infrared temperature sensor, and vibration signal sequence data collected by the vibration sensor, which are used to reflect the dynamic operating status of the tension clamp, such as temperature change trends and mechanical vibration patterns. The environmental data includes temperature, humidity, wind speed, and rainfall collected by environmental sensors, which are used to reflect the impact of the external environment on the tension clamp. The preprocessing and feature extraction of the two types of data need to be adapted to their data type characteristics respectively.

[0067] In some implementations, the preprocessing and feature extraction of time-series and environmental data may include the following steps: For time-series data, the temperature time-series data and vibration signal sequences are first preprocessed, including using Kalman filtering or wavelet thresholding to remove redundant information such as equipment noise and electromagnetic interference. Next, a one-dimensional convolutional network (1D-CNN) is used to extract temporal features. By sliding convolutional kernels, the local variation patterns of the data are captured. The output features include, but are not limited to, heating rate (unit: ℃ / min), temperature fluctuation pattern, and vibration frequency domain energy (unit: dB). Then, the extracted temporal features are integrated into a fixed-dimensional vector, namely the second feature vector.

[0068] Meanwhile, to quantify the reliability of time series data, a signal-to-noise ratio (SNR) score needs to be calculated. The formula is SNR = effective signal amplitude squared / noise signal amplitude squared. The effective signal amplitude refers to the signal amplitude in the time series data that reflects the actual state change, such as the amplitude of normal temperature rise and fall. The noise signal amplitude refers to the amplitude of the interference signal that has been removed. The higher the SNR, the stronger the data reliability.

[0069] For environmental data, firstly, the environmental data is preprocessed, including filling in missing values ​​and removing outliers; then, the stability index of the environmental data is calculated. First, the degree of data fluctuation is quantified by the formula fluctuation coefficient = standard deviation / mean. Then, the degree of fluctuation is converted into a stability index by the stability score = 1 - fluctuation coefficient. That is, the smaller the fluctuation coefficient, the closer the stability score is to 1, and the more stable the data. Finally, the environmental data and stability score are input into the fully connected layer, and embedded representation is performed through linear transformation and activation function to output a fixed-dimensional vector, namely the third feature vector.

[0070] It should be noted that the core value of time-series characteristics lies in capturing dynamic anomalies, such as sudden temperature rises and abrupt changes in vibration frequency, providing a basis for overheating and degradation warnings; the core value of environmental characteristics lies in distinguishing between equipment malfunctions and environmental influences, such as temperature rise caused by hot weather vs. temperature rise caused by poor contact of the clamp. Combining the two can avoid misjudgments caused by ignoring environmental factors.

[0071] S203. Fuse multiple types of feature vectors to generate the final multimodal feature vector.

[0072] In some implementations, obtaining multimodal feature vectors includes the following steps: The first step is feature dimension alignment. The dimensions of the first, second, and third feature vectors are unified through a fully connected layer to avoid fusion bias caused by dimensional differences. The second step is to weight the feature confidence scores. The image sharpness score of S201, the temporal signal-to-noise ratio score of S202, and the environmental stability score are mapped to the [0,1] interval through Min-Max normalization, and used as the weights of the corresponding feature vectors to weight the dimension-aligned features. The third step is feature concatenation and integration, which involves concatenating the weighted feature vectors of the three types in dimensional order to generate the final multimodal feature vector.

[0073] It should be noted that dimensional alignment is the foundation of fusion, which can avoid the problem of high-dimensional features masking low-dimensional features; confidence weighting is the key to fusion, which can give higher weight to more reliable features in the final vector, thereby improving the overall effectiveness of the features; the combination of the two can ensure that the multimodal feature vector integrates full-dimensional information and highlights the role of high-value data.

[0074] For example, the 128-dimensional first feature vector of S201, the 128-dimensional second feature vector of S202, and the 64-dimensional third feature vector are first unified into 128 dimensions through a fully connected layer; then the image sharpness score of 0.82, the temporal signal-to-noise ratio score of 0.95, and the environmental stability score of 0.946 are normalized by Min-Max (assuming that the minimum value of the three scores is 0.5 and the maximum value is 1.0), resulting in normalized confidence scores of 0.64, 0.9, and 0.892, respectively; after weighting the three types of 128-dimensional features (first × 0.64, second × 0.9, third × 0.892), they are concatenated in the order of "first-second-third" to generate the final 384-dimensional multimodal feature vector.

[0075] Based on the above technical solution, S2 achieves efficient transformation and value mining of heterogeneous data through categorized preprocessing, targeted feature extraction, and confidence-weighted fusion. For example, it focuses on enhancing structural details in X-ray images, capturing dynamic trends in time-series data, and quantifying influencing factors in environmental data. Finally, it generates standardized multimodal feature vectors through dimensional alignment and confidence-weighted fusion. This process not only solves the problems of difficult data fusion and low feature effectiveness in traditional methods, but also provides high-quality input for subsequent multimodal interactive guidance networks, laying a core foundation for improving early warning accuracy.

[0076] In one possible implementation of the embodiments of this application, combined with Figure 2 ,like Figure 4 As shown, the multimodal interaction guidance network in S3 is the core intelligent decision-making unit for multi-category early warning tasks of tension clamps. Its design aims to solve the challenges of heterogeneous multimodal data fusion and multi-task collaborative optimization. The overall framework consists of three parts: a shared feature extraction module, a task interaction guidance module, and an early warning output module. These three modules form a serial collaborative architecture. Specifically, the shared feature extraction module is used to achieve unified preprocessing of multimodal data; the task interaction guidance module is used to handle deep interaction of multi-task features; and the early warning output module is used to achieve quantitative output of risk probabilities. This ultimately forms an end-to-end processing flow from multimodal feature input to multi-category early warning probability output, which can be implemented through the following S301, S302, and S303, as explained in detail below: S301. The multimodal feature vectors are processed by the shared feature extraction module to generate shared feature vectors.

[0077] Among them, the shared feature extraction module is the basic fusion layer of the multimodal interaction guidance network. Its core function is to solve the problems of heterogeneity, redundancy and weak spatiotemporal correlation of multimodal features. Through cross-modal alignment and feature optimization, it transforms the scattered multimodal feature vectors into shared feature vectors with unified semantics and scale, providing a consistent input basis for subsequent multi-task interactions.

[0078] In some implementations, the shared feature extraction module consists of two levels of sub-units: a cross-module fusion unit and a feature enhancement unit. The specific workflow is as follows: The shared feature extraction module is the data preprocessing hub of the network. Its core function is to receive the multimodal feature vectors generated by S2, and through cross-modal alignment, weighted fusion, and feature enhancement, eliminate dimensional differences and redundant interference from heterogeneous data to generate shared feature vectors with unified semantics and scale, providing consistent input for subsequent multi-task interactions. This module consists of two levels of sub-units: a cross-modal fusion unit and a feature enhancement unit. The specific workflow is as follows: 1. Cross-module fusion unit: Alignment and weighting of multimodal features.

[0079] The core objective of the cross-module fusion unit is to address the issues of inconsistent multimodal feature dimensions and weak correlations. It achieves data fusion through three steps: spatiotemporal correlation calculation, dimension alignment, and confidence weighting. The first step is to calculate the spatiotemporal correlation: a multi-head attention mechanism is adopted to decompose the multimodal feature vector into image modality branch, temporal modality branch and environment modality branch. The spatiotemporal correlation strength between any two modality branches is calculated by an 8-head attention structure, and a 3×3 modality correlation weight matrix is ​​output to quantify the collaborative value of different modal data. Among them, image-time correlation reflects the coupling relationship between structural defects and temperature changes, time-environment correlation reflects the dependence of temperature fluctuations on external temperature and humidity, and image-environment correlation reflects the coupling influence between the internal structural state of tension clamps and external environmental factors. For example, when there are crimping gaps inside the clamp, high humidity and high rainfall will accelerate metal corrosion at the gaps, leading to further expansion of structural defects. Strong winds may cause uneven stress on clamps with potential strand breakage, exacerbating the risk of strand breakage. This correlation can quantify the promoting or inhibiting effect of the external environment on the development of structural defects, providing a basis for subsequent fusion of multimodal features to judge the evolution trend of defects. The second step is modal feature dimension alignment: To address the dimensional differences among the modal branches, an adaptive fully connected layer is used to map the environmental modal features from 64 dimensions to 128 dimensions, ensuring that the dimensions of the three modal branches are consistent and avoiding fusion bias caused by dimensional differences. At the same time, the fully connected layer introduces nonlinear transformation through the ReLU activation function to enhance the expressive power of the modal features.

[0080] The third step, confidence-weighted fusion: Introducing the modal confidences calculated in S2, and combining them with the modal relevance weights from the first step, the dimension-aligned features are weighted and fused. The fusion formula is as follows: Cross-modal fusion feature = image feature × image confidence × image-global association weight + temporal feature × temporal confidence × temporal-global association weight + environmental feature × environmental confidence × environmental-global association weight; where the modality-global association weight is the average of the association weights of the corresponding modality with all other modalities in the modality correlation matrix, ensuring that high-confidence and high-correlation modality features dominate the fusion result and improve the effectiveness of the fusion features.

[0081] 2. Feature enhancement unit: Noise reduction and scale optimization of fused features.

[0082] The feature enhancement unit addresses the issues of noise interference and gradient vanishing in fused features, achieving feature optimization through residual connections, batch normalization, and scale unification. Residual Connection: Sets a shortcut path to directly add cross-modal fused features to the subsequent processing results, constructing a residual mapping between input and output, effectively alleviating the gradient vanishing problem in deep learning model training, and enhancing the preservation of the original semantic information of fused features; Batch Normalization: Standardizes the fused features after residual connection, adjusts the feature value distribution to a normal distribution with mean 0 and variance 1, eliminates noise interference caused by differences in the original distribution of features of different modalities, and improves the stability of model training. Scale uniformity: The batch normalized fused features are compressed to 64 dimensions through a fully connected layer, outputting a shared feature vector with uniform dimensions, low noise, and strong semantics, thus completing the processing flow of the shared feature extraction module.

[0083] It should be noted that the core value of the shared feature extraction module lies in the dual optimization of multi-head attention and confidence weighting: multi-head attention breaks through the limitation of traditional linear fusion in capturing deep correlations, while confidence weighting solves the problem of low-quality modalities dragging down the overall fusion effect. The combination of the two makes the shared feature vector both comprehensive and accurate.

[0084] S302. Calculate the task relevance weights through the task attention gate of the task interaction guidance module to generate the initial task feature vector.

[0085] Among them, the task attention gate is a feature filtering layer for multi-task interaction. Its core function is to assign differentiated weights to the shared feature vectors based on the overlap of feature requirements of multiple categories of early warning tasks, so that each task can prioritize the acquisition of feature information with high relevance to it, and avoid the problem of insufficient task adaptability caused by indiscriminate use of shared features.

[0086] In some implementations, the workflow of the task attention gate is as follows: First, for multiple types of early warning tasks, independent fully connected layers are set up to map the shared feature vectors into exclusive query vectors that match the task requirements. For example, the query vector for defect early warning focuses on the structural edge feature dimension, and the query vector for fever early warning focuses on the temperature change feature dimension. The dimension of each query vector is consistent with the shared feature vector to ensure the feasibility of subsequent correlation calculation.

[0087] Next, cosine similarity is used to calculate the correlation between any two task-specific query vectors, quantifying the overlap of feature requirements between tasks. The calculation formula is: cosθ=(Q_i·Q_j) / (||Q_i||×||Q_j||); where Q_i and Q_j are the task-specific query vectors of the i-th and j-th classes, respectively. " represents the vector dot product, and ||Q_i|| and ||Q_j|| represent the L2 norm of the vector; an n×n task relevance matrix is ​​generated using this formula, where n represents the number of early warning task categories.

[0088] Then, Softmax normalization is performed on each row of the task relevance matrix to convert the relevance into attention weights that can be directly applied. After normalization, an n×n task attention weight matrix is ​​obtained, with the weights of each row summing to 1. Based on this matrix, the shared feature vectors are weighted and summed to generate the initial task feature vectors for each of the multiple task classes.

[0089] It should be noted that the core of cosine similarity is to quantify the feature association between tasks, while the core of Softmax is to transform the association into assignable weights. The combination of the two ensures that the features of highly correlated tasks can be preferentially utilized by the current task.

[0090] S303. Cross-task feature transfer and fusion are achieved through the interactive task guide module of the task interaction guide module, generating task-specific feature vectors.

[0091] Among them, the interactive task facilitator is a deep fusion layer for multi-task collaboration. Its core function is to correct the features of the preceding task and cross-task features through feature transfer, so that the features of each task simultaneously contain its own core information and other task auxiliary information, breaking the limitations of shallow feature sharing and weak mutual guidance in traditional multi-task learning.

[0092] In some implementations, the interactive task facilitator is based on a series of gated loop units, and its specific workflow may include: The first step is to sort the four initial task feature vectors generated by S302 according to the preset logic of "defect warning → heat warning → corrosion warning → degradation warning". The sorting is based on the causal relationship between tasks, such as defects may cause heat, heat may accelerate corrosion, and corrosion may lead to degradation. The sorted initial feature vectors are then sequentially input into a series of gated recurrent units (GRUs). The GRU uses update gates and reset gates to take the initial features of the preceding tasks as the prior inputs of the following tasks, realizing the temporal transmission and dynamic correction of features between tasks, and finally obtaining the guiding intermediate features of each task.

[0093] The second step is to perform feature cross-concatenation on the intermediate features of each task: concatenate the intermediate features of the current task with the intermediate features of the other three tasks by dimension, and integrate cross-task feature information to avoid feature isolation between tasks.

[0094] The third step is to input the cross-stitched high-dimensional features into the fully connected layer for dimensionality compression, and at the same time combine the GELU activation function (Gaussian Error Linear Units) to enhance feature representation: GELU is activated by Gaussian distribution probability weighting, which can more accurately capture subtle differences in features; The final output is a specific feature vector for each task, which is the core output of the multimodal interactive guidance network for each early warning task.

[0095] It should be noted that the core value of GRU is to enable effective features of preceding tasks to accurately empower subsequent tasks. For example, defect features help fever warning distinguish between "fever caused by defects" and "fever caused by the environment". Feature cross-concatenation forces information complementarity between tasks. The combination of the two allows the specific feature vector of each task to focus on itself while taking into account the whole.

[0096] Based on the above technical solutions, S3 constructs a deep correlation mechanism between multimodal data and multiple early warning tasks through shared feature extraction, task attention screening, and interactive guidance fusion. Shared feature extraction solves the unification problem of heterogeneous data, task attention gates realize task-adaptive feature screening, and interactive task guides complete dynamic feature complementarity between tasks. This not only taps into the deep value of multimodal data but also transforms early warning tasks such as defects, overheating, corrosion, and degradation from independent learning to collaborative optimization. It can improve feature sharing efficiency and task guidance accuracy, providing high-quality task-specific feature support for S4 to calculate accurate early warning probabilities and achieve hierarchical decision-making.

[0097] In one possible implementation of the embodiments of this application, combined with Figure 2 ,like Figure 5 As shown, the above S4 specifically includes the following S401 to S403: S401. Process task-specific feature vectors through the early warning classification head to generate initial probabilities for multi-category early warning tasks.

[0098] Among them, the early warning classification head is the probability output layer of the multimodal interactive guidance network. Each early warning task corresponds to an independent early warning classification head, which transforms the task-specific feature vector generated by S3 into a quantifiable initial probability of early warning, intuitively reflecting the possibility of a corresponding fault in the tension clamp.

[0099] In some implementations, S401 may include the following steps: Each early warning classification head receives a specific feature vector corresponding to the task; By using a fully connected layer to compress the dimension and enhance the features of specific feature vectors, a 64-dimensional vector is mapped to a 1-dimensional feature. At the same time, a non-linear transformation is introduced by combining the GELU activation function to dynamically adjust the activation intensity according to the probability distribution of the feature values, so as to more accurately capture the subtle risk differences in the features. Then, the Dropout layer is used for regularization, randomly discarding some neurons to avoid the model from causing a bias in the warning probability due to overfitting the training data. Finally, the processed feature values ​​are mapped to the [0,1] interval using the Sigmoid activation function to obtain the initial probability of the multi-class early warning task. The closer the value is to 1, the higher the probability of the corresponding fault occurring.

[0100] It should be noted that the design of the independent warning classification head ensures that the probability calculation of each task is not interfered with by the features of other tasks, and the combination of GELU and Dropout balances the feature representation accuracy and the model generalization ability, avoiding probability distortion caused by single activation or lack of regularization.

[0101] S402. Generate the initial probability of the security state category through the security classification header and reverse association logic.

[0102] Among them, the safety classification head is a key module to supplement the "non-fault state" assessment. By quantifying the comprehensive abnormality of multiple types of early warning tasks, it reverse-derives the confidence level of the tension clamp being in a safe state and obtains the initial probability of the safe state category.

[0103] In some implementations, obtaining the initial probability of a security state category may include the following steps: The first step is for the safety classification head to receive the initial probabilities of all warning tasks output by S401, and to calculate the comprehensive anomaly index using a weighted summation formula based on reverse correlation logic: Anomaly Index = Σ (Initial Probability of Warning Task × Task Weight), where the task weight is set according to the severity of the fault. For example, defects and overheating faults are more dangerous, so the weight is set to 0.3; corrosion and degradation are set to 0.2, ensuring that the sum of the weights is 1. This formula reflects the difference in contribution of different faults to the overall anomaly through the weights. The second step is to perform a reverse mapping on the anomaly degree index to calculate the basic safety confidence value. The formula is: Basic safety confidence value = 1 - anomaly degree index, which directly reflects the negative correlation between the higher the anomaly degree and the lower the safety confidence value. The third step is to input the basic security confidence values ​​into the fully connected layer for feature optimization, and then map them to the [0,1] interval through the Sigmoid activation function to obtain the initial probability of the security state category.

[0104] It should be noted that the core value of reverse correlation logic lies in global risk balancing, avoiding misjudging the overall state when the probability of a single failure is high but the probability of other failures is low; at the same time, the supplement of the probability of safe state makes the evaluation dimensions more complete.

[0105] S403. Normalize the probabilities of all categories and execute hierarchical early warning decisions based on the normalized probabilities.

[0106] In some implementations, tiered early warning decision-making may include the following steps: First, the initial probabilities of the four warnings of S401 and the initial probability of safety of S402 are normalized by the Softmax function. After normalization, the sum of the probabilities of each category is 1, forming a probability distribution of "safety-defect-heating-corrosion-degradation".

[0107] Then, based on the normalized probability, a tiered threshold is set, and an early warning decision is made. Finally, the corresponding level is matched based on the highest value of the normalized probability of each category, and the warning results and operation and maintenance suggestions are output.

[0108] For example, a tiered early warning decision-making process can be implemented as follows: For the safety category, if the normalized probability is greater than 0.85, the status is determined to be safe. For multi-category early warning tasks, if the normalized probability of a certain early warning task is >0.9, it indicates that the probability of that category is extremely high, and it is judged as a level 1 early warning. If the normalized probability of a certain early warning task is ∈ (0.85, 0.9], it is determined to be a level 2 early warning. If the normalized probability of a certain early warning task is ∈ (0.7, 0.85], then it is determined to be a level three early warning. If the normalization of a certain early warning task is ∈(0.6,0.7], it indicates that there are signs of potential dangers and is judged as a level four early warning.

[0109] Based on the above technical solution, S4 achieves the quantitative accuracy of early warning probability, the completeness of security status assessment, the practicality of early warning decisions, and the fine-grained identification and graded response of multiple types of risks through a three-level collaborative process of task-specific initial probability calculation, reverse derivation of security probability, probability normalization, and graded early warning decision-making. It utilizes the targeted feature mapping of independent early warning classification heads, the global anomaly quantification of reverse correlation logic, the probability distribution unification of the Softmax function, and the risk response adaptation mechanism of graded thresholds. This solves the technical problems in existing technologies, such as isolated and unrelated early warning probabilities of multiple types, monitoring blind spots caused by the lack of security status assessment, frequent false alarms and missed alarms caused by the lack of quantitative basis for early warning decisions, and the roughness and waste of operation and maintenance costs caused by risk response relying on only a single threshold.

[0110] Although this application has been described herein in conjunction with various embodiments, those skilled in the art, by reviewing the accompanying drawings, disclosure, and appended claims, will understand and implement other variations of the disclosed embodiments in carrying out the claimed application. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude multiple instances. A single processor or other unit can implement several functions listed in the claims. While different dependent claims may recite certain measures, this does not mean that these measures cannot be combined to produce good results.

[0111] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and drawings are merely exemplary illustrations of this application as defined by the appended claims, and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from the spirit and scope of this application. Thus, if such modifications and modifications of this application fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and modifications.

Claims

1. A state early warning method for a strain clamp of a power transmission line, characterized in that, The application relates to a multi-modal interactive guidance network for early warning of a strain clamp, and belongs to the technical field of early warning of a strain clamp. Real-time data of a strain clamp is collected through a multi-modal sensor array; the multi-modal sensor array comprises a thermal resonator, an X-ray imaging device, an infrared temperature sensor, a vibration sensor and an environmental sensor, and the thermal resonator comprises a resonant cavity, an excitation unit, a detection unit and a signal processing circuit; The collected real-time data is preprocessed and feature-extracted to generate a multi-modal feature vector; The multi-modal feature vector is input into a multi-modal interactive guidance network constructed based on a deep learning algorithm, and mutual guidance and feature sharing between multi-class early warning tasks are performed through a task interactive guidance module; Based on the output of the multi-modal interactive guidance network, the early warning probability of the multi-class task is calculated, and a hierarchical early warning decision is made.

2. The method for early warning of the state of a strain clamp for a power transmission line according to claim 1, characterized in that, The thermal resonator works based on a temperature-frequency conversion mechanism and comprises the following steps: The surface temperature of the strain clamp is sensed through a thermal material in the resonant cavity, and the frequency offset generated by the surface temperature change is obtained through the excitation unit and the detection unit; converting the frequency offset into a temperature value based on a pre-calibrated thermal model, the thermal model being: Δf=k x ΔT wherein Δf is the frequency offset, ΔT is the temperature change, and k is a thermal coefficient, is an environmental compensation term. The processed temperature value is sent to an edge computing node through a wireless transmission module.

3. The method for early warning of the state of a strain clamp for a power transmission line according to claim 1, characterized in that, The real-time data comprises: surface temperature time series data of the strain clamp collected through the thermal resonator; X-ray image data collected through the X-ray imaging device; temperature time series data collected through the infrared temperature sensor; vibration signal sequence data collected through the vibration sensor; environmental data of temperature, humidity, wind speed and rainfall collected through the environmental sensor.

4. The method for early warning of the state of a strain clamp for a power transmission line according to claim 3, characterized in that, The preprocessing and feature extraction of the collected real-time data to generate the multi-modal feature vector comprise the following steps: X-ray images are enhanced through a Retinex algorithm, and super-resolution reconstruction is performed through a generative adversarial network to output a first feature vector; one-dimensional convolution network is used to extract time sequence features from the temperature time series data and the vibration signal sequence data to obtain a second feature vector; the time sequence features comprise a temperature rising rate, a fluctuation mode and frequency energy; environmental data is embedded through a full connection layer to obtain a third feature vector; the first feature vector, the second feature vector and the third feature vector are spliced to obtain a fused multi-modal feature vector.

5. The method for early warning of the state of a strain clamp for a power transmission line according to claim 1, characterized in that, The multi-modal interactive guidance network comprises a shared feature extraction module, a task interactive guidance module and an early warning output module; wherein, the shared feature extraction module comprises a cross-modal fusion unit and a feature enhancement unit, and is used for cross-modal alignment and feature optimization of the multi-modal feature vector to obtain a shared feature vector; the cross-modal alignment means that the spatio-temporal correlation between different modal features is calculated through a multi-head attention mechanism, and the feature optimization means that noise reduction and scale unification are performed on the fused features through a residual connection and a batch normalization technology, and a higher weight is given to the modal features with higher confidence through an adaptive weight adjustment mechanism; The task interaction guiding module includes a task attention gate and an interaction task guide, and is configured to perform weight distribution and feature guidance among the shared feature vectors to obtain specific feature vectors of each task; the weight distribution represents calculating attention weights based on correlations among the tasks, and the feature guidance represents performing transmission and correction of features among the tasks through a gated recurrent unit, taking feature vectors output by different tasks as prior inputs of the remaining tasks, and cross-guiding feature sharing and learning among the tasks; The pre-warning output module includes a plurality of pre-warning classification heads, and is configured to perform multi-class probability calculation on the specific feature vectors of each task to obtain pre-warning probabilities of multi-class tasks.

6. The method for early warning of the state of a strain clamp for a power transmission line according to claim 5, characterized in that, An internal working process of the shared feature extraction module is as follows: The cross-modal fusion unit receives multi-modal feature vectors, calculates correlation weights among different modalities through a multi-head attention mechanism, and performs dimension alignment on the feature vectors of each modality through an adaptive fully connected layer to obtain the dimension-aligned feature vectors of each modality; Based on the correlation weights and the confidence of each modality, the dimension-aligned feature vectors of each modality are weighted and fused to obtain cross-modal fusion features; The feature enhancement unit is configured to input the cross-modal fusion features, strengthen feature expression through a residual connection structure, and perform noise reduction processing on the fusion features through a batch normalization technique to obtain noise-reduced fusion features; The fully connected layer is configured to perform feature scale unification on the noise-reduced fusion features to output a shared feature vector.

7. The method for early warning of the state of a strain clamp for a power transmission line according to claim 5, characterized in that, An internal working process of the task interaction guiding module is as follows: The shared feature vector is received, and an independent fully connected layer in the task attention gate is used to generate a dedicated query vector for each multi-class pre-warning task; The correlation of each multi-class is calculated based on the dedicated query vector through a cosine similarity, and the calculation result is subjected to Softmax normalization processing to obtain a task attention weight matrix; Based on the task attention weight matrix, the shared feature vector is weighted and summed to obtain an initial task feature vector of each task; The initial task feature vector is input into the interaction task guide after being sorted according to a preset logic, and the feature vectors of the previous tasks are taken as prior inputs of the subsequent tasks through the series of gated recurrent units to perform transmission and correction of task key features, thereby obtaining guided intermediate features of each task; For each guided feature, the guided intermediate feature of itself and the guided intermediate features of other tasks are fused through a feature cross-splicing operation, and then the dimension is compressed through a fully connected layer and the feature expression is strengthened through a GELU activation function to output a specific feature vector of each task.

8. The method for early warning of the state of a strain clamp for a power transmission line according to claim 5, characterized in that, An internal working process of the pre-warning output module is as follows: Each pre-warning classification head receives the specific feature vector of each task output by the task interaction guiding module through a fully connected layer and a GELU activation function to strengthen and compress the dimension of the input specific feature vector, performs regularization through a Dropout layer, and then calculates an initial probability of the corresponding pre-warning task through a Sigmoid activation function. ​ The initial probability of the multi-category early warning task is received by the security classification head, and the initial probability of the security state category is calculated based on the reverse correlation logic through a fully connected layer and a Sigmoid activation function. The initial probabilities of all category early warnings and the initial probability of the security state category are subjected to Softmax normalization processing, so that the sum of the probabilities of each category is 1, and the probability distribution of the multi-category early warning task and the security state category is finally output.

9. The method for early warning of the state of a strain clamp for a power transmission line according to claim 8, characterized in that, The reverse correlation logic refers to a negative correlation mapping relationship between the initial probability of the security state category and the initial probability of each early warning task, and the confidence degree of the strain clamp in the security state category is determined by quantifying the comprehensive abnormality degree of the multi-category early warning task, and the operation steps include: The initial probability of the multi-category early warning task is comprehensively calculated, and an abnormality degree index is obtained through a weighted summation formula; The abnormality degree index is subjected to reverse mapping, and a basic security confidence value is obtained through linear transformation of "1-abnormality degree index"; The basic security confidence value is sequentially input into a fully connected layer and a Sigmoid function to obtain the initial probability of the security state category.

10. A method for early warning of the state of a strain clamp for a power transmission line, characterized in that, It includes: a data acquisition module, an intelligent processing module and an early warning decision module; wherein The data acquisition module is configured to acquire real-time data of the strain clamp through a multi-modal sensor array, the multi-modal sensor array includes a thermal resonator, an X-ray imaging device, an infrared temperature sensor, a vibration sensor and an environmental sensor, wherein the thermal resonator includes a resonant cavity, an excitation unit, a detection unit and a signal processing circuit; The intelligent processing module is configured to pre-process and extract features from the acquired real-time data, generate a multi-modal feature vector, and input the multi-modal feature vector into a multi-modal interaction guiding network constructed based on a deep learning algorithm, and guide and share features between multi-category early warning tasks through a task interaction guiding module; The early warning decision module is configured to calculate the early warning probability of the multi-category task based on the output of the multi-modal interaction guiding network and make a hierarchical early warning decision.

Images